Elucidation of complex diseases with machine learning

Rule-network (RN) construction

Interdependencies are discovered from transparent classifiers, such as rule-based classifiers (RBCs).

The conjuncts are co-predictive of the decision.

Nodes - conjuncts in rules, i.e. features with their values occurring in the rules.

Edge - connects two nodes if the corresponding conjuncts occur in the same rule.

RNs are built for each decision (outcome class).

VisuNet

• RN visualization tool implemented as an R package,
• helps identifying the strongest features and their interdependencies,
• easy to use due to an interactive interface,
• applicable for any RBC,
• publicly available on GitHub: https://github.com/komorowskilab/VisuNet

Applications

MG - Mateusz Garbulowski

FB&ZK - Fredrik Barrenäs & Zeeshan Khaliq

KD - Klev Diamanti

KS - Karolina Smolinska

Case-control studies of autism - sample RN

4 genes were previously reported in the SFARI Gene database of autism associations: COX2, TMLHE, NCKAP5L and TSPOAP1.

Several genes identified in the literature as autism associated: ZSCAN18, SCIN, NCS1, CAPS2, ZFP36L2.

MAP7 and COX2 interdependency in autism

Project 1 - Summary

• VisuNet is conceptually simple, but useful for interpreting RBCs,
• RN identifies different combinations of features and their levels that are co-predictors,
• RN helps explain the differences between the decision classes,
• autism-control RN confirms genes associations with autism,
• RN may help to discover the interaction mechanisms for attributes, e.g. MAP7 and CAPS2.

The manuscript is planned to finish by the end of this year.

Project 2: funMotifs and the functional annotation of the tissue-specific noncoding mutations

The gene expression is regulated by transcription factors (TFs) through binding to TF biding sites (motifs).
To understand how the genes are regulated, we need to know what determinated functionality of TF motifs.

We incorporated data from large-scale genomics platforms including ENCODE, RoadMap Epigenomics and FANTOM to characterize the tissue-specific functional TF motifs.

As a consequence, funMotifs was developed by Husen M. Umer. I was mainly working on the framework applications.

Two manuscripts were written based on the Project 2.

Manuscript 1

funMotifs framework

CFRMs distribution

Applications

HS - Husen M. Umer

KS - Karolina Smolinska

Comparison to existing annotation databases

2,219,330 eQTL SNPs were annotated using funMotifs, RegulomeDB and HaploReg.

RegulomeDB: Boyle,A.P., et al. (2012)

HaploReg: Ward,L.D., et al. (2016)

Overview of the candidate functional GWAS SNP rs8103622 in breast tissue

Manuscript 2

The aim of the study

Functional annotation of mutations in the Pan Cancer Analysis of Whole Genomes consortium data for 2,500 cancer genomes in 37 cancer types.

Results

Using the funMotifs framework we identified 121,029 CFRMs overlapped 217,075 potentially functional motifs across the 44 cohorts. Many CFRMs were found in genes that were previously linked with cancer.

Overview of mutations

Workflow

CEBPB - the most significantly mutated motifs

The active motifs of CEBPB are significantly less methylated genome-wide than inactive motifs and still there is an abundance of mutations at this site (P-value < 1e-50).

Highly recurrent mutated regulatory elements

Project 2 - Summary

• Characterization of the noncoding regions of human genome is important to understand mechanisms that lead to complex diseases,
• Functional annotations help to identify regulatory mutations that may significantly dysregulate gene expression in cancer and other complex diseases,
• funMotifs applications identified potential functional variants in cancer and normal genomes,
• Experimental validation may be performed on the CFRMs to confirm our findings

funMotifs was implemented as a command-line pipeline (https://github.com/komorowskilab/funMotifs) and a web-based interface: http://bioinf.icm.uu.se/funmotifs.

Future plans

• Project 1. - implement the possibility to annotate the RN attributes using biological properties of the input datatypes and various databases, e.g. GO terms for a gene expression data.

• Project 2. - modification of the funMotifs algorithm to improve identification of the false positive regulatory motifs. We are planning to continue research on the functional annotation of motifs. We would like to extend the functional motifs definition in terms of the conservation of the individual nucleotides in TF motifs across various diseases.

• Project 3. - identification and visualization of interdependent genes that discern patients in the case-control studies of autism spectrum disorders by combining multiple rule-based classifiers.

• Project 4. - identification of potential biomarkers that discern glioma subtypes using CFRMs from the COSMIC-available mutations for gliomas. Intersection of CFRMs with the differentially expressed genes based on the TCGA and GTEx databases.

Publications:

1. Husen M. Umer, Karolina Smolinska-Garbulowska, Nour-al-dain Marzouka, Zeeshan Khaliq, Claes Wadelius, Jan Komorowski, funMotifs: Tissue-specific transcription factor motifs, 2019, bioRxiv 683722; doi: https://doi.org/10.1101/683722 (subbmited)

2. Husen M. Umer, Karolina Smolinska-Garbulowska, Jan Komorowski, Claes Wadelius, Functional annotation of noncoding mutations in cancer, 2019(manuscript)

3. Marco Cavalli, Nicholas Baltzer, Gang Pan, José Ramón Bárcenas Walls, Karolina Smolinska Garbulowska, Chanchal Kumar, Stanko Skrtic, Jan Komorowski, and Claes Wadelius. Studies of liver tissue identifies functional gene regulatory elements associated to gene expression, type 2 diabetes and other metabolic diseases, Human genomics 13, no. 1 (2019): 20

4. Mateusz Garbulowski, Klev Diamanti, Karolina Smolinska-Garbulowska, Patricia Stoll, Susanne Bornelöv, Aleksander Øhrn, Jan Komorowski, (2019), R.ROSETTA: a package for analysis of rule-based classification models, bioRxiv 625905; doi: https://doi.org/10.1101/625905 These authors contributed equally to the work as second authors.

Posters:

1. Michal J. Dabrowski, Michal Draminski, Klev Diamanti, Mateusz Garbulowski, Karolina Smolinska, Jacek Koronacki, Jan Komorowski. Identification of the boundary between common and ancestry-dependent levels of gene expression of Homo sapiens adaptive immune system. Systems Biology: Networks, Cold Spring Harbor Laboratory, Cold Spring Harbor, USA 14-18 March 2017.
2. Mateusz Garbulowski, Karolina Smolinska, Klev Diamanti, Patricia Stoll, Zeeshan Khaliq, Sara Younes, Behrooz Torabi, Susanne Bornelöv, Jan Komorowski. RROSETTA: a toolkit for creating rough set models. Swedish Bioinformatics Workshop 2017, Uppsala, Sweden, 13-14 November 2017.
3. Mateusz Garbulowski, Karolina Smolinska, Klev Diamanti, Patricia Stoll, Zeeshan Khaliq, Sara Younes, Behrooz Torabi, Susanne Bornelöv, Jan Komorowski. RROSETTA: a toolkit for creating rough set models. Swedish Epigenetics and Chromatin Meeting, EpiChrom 2017, Uppsala, Sweden, 30 November - 1 December 2017.
4. Karolina Smolinska-Garbulowska, Steven O. O. Anyango, Jan Komorowski, VisuNet: an interactive tool for rule network visualization of rule-based models, SciLifeLab Science Summit 2019, Artificial Intelligence for Life Sciences, Uppsala, May 15, 2019

Presentation:

1. Karolina Smolinska-Garbulowska, VisuNet: an interactive tool for rule network visualization of rule-based models, the awarded Flash Talk, SciLifeLab Science Summit 2019, Artificial Intelligence for Life Sciences, Uppsala, May 15, 2019

Acknowledgements